fix(a2a-demo): adapt joint graph to ADK 1.33 RemoteA2aAgent sub-session telemetry

[caohy1988]

Live-verified on test-project-0728-467323 with google-adk==1.33.0.

This PR fixes the ADK 1.33 RemoteA2aAgent telemetry shape and removes the receiver-extraction flake that was blocking the A2A demo from running end-to-end.

What changed

ADK 1.33 caller-side sub-session mapping

ADK 1.33 writes RemoteA2aAgent telemetry under a fresh caller-side sub-session (agent='audience_risk_reviewer'), not under the parent supervisor session. The demo now materializes caller.supervisor_a2a_invocations to bridge that split:

• Supervisor side: current-run TOOL_STARTING rows for audience_risk_reviewer.
• A2A side: A2A_INTERACTION rows from the RemoteA2aAgent sub-sessions.
• Pairing: first A2A row inside the tool-call time window (supervisor_ts <= a2a_ts < next_supervisor_ts) for the same user_id.

Gates now match that shape:

• G1: dataset-level count of A2A_INTERACTION rows for agent='audience_risk_reviewer'.
• G1.5: mapping count == campaign count and every row has non-null a2a_context_id.
• G3: receiver stitch reads a2a_context_id from the mapping table.

build_joint_graph.remote_agent_invocations reads from the mapping. The graph DDL is unchanged: caller_session_id still points at the supervisor session, while a2a_context_id still stitches to the receiver session.

Receiver extraction reliability

The SDK decision extraction query now follows current BigQuery AI.GENERATE docs: it uses SQL-style output_schema for the nested decision/candidate contract and converts AI.GENERATE(...).decisions back to JSON with TO_JSON_STRING, preserving the existing Python parser contract.

build_org_graphs.py also has a demo-specific safety net: if AI.GENERATE returns too few receiver decisions/candidates, it deterministically parses the receiver agent's strict prompt-shaped LLM_RESPONSE format and rewrites decision_points / candidates, then recreates decision edges and reruns the gate.

Live verification

• G1/G1.5/G2/G3 green with ADK 1.33 telemetry.
• Time-window mapping returns exactly 3 supervisor→A2A rows.
• Receiver extraction initially showed decision_points=1, candidates=3; fallback repaired it to decision_points=4, candidates=12, then the acceptance gate passed.
• build_joint_graph.py after repair:
  • stitch coverage a2a_calls=3, stitched_edges=3
  • scoped receiver coverage decision_points=3, candidates=9
  • per-session receiver gate passed
  • traversal smoke returned rows across campaign → remote invocation → receiver → decision → option
• run_analyst_agent.py completed all four canned questions and produced user-visible answers for graph health, campaign list, one campaign audit path, and portfolio dropped-option review.

Static verification

• pytest tests/test_context_graph.py -q → 98 passed
• pyink --check on changed Python → clean
• isort --check-only on changed Python → clean
• py_compile on changed Python → clean
• bash -n on demo shell scripts → clean
• git diff --check → clean

Reference

BigQuery AI.GENERATE supports structured output using SQL-style output_schema and returns a STRUCT; when output schema is specified, result is replaced by the custom schema fields. That is why the new query reads .decisions and wraps it with TO_JSON_STRING for the existing parser.

Docs: https://docs.cloud.google.com/bigquery/docs/reference/standard-sql/bigqueryml-syntax-ai-generate